@HelloIntento

Enabling global companies to translate 20x more on the same budget, making their customer and employee experience instantly multilingual

 San Francisco, CA
inten.to
Joined March 2016
  • Tweets646
  • Following32
  • Followers262
The 2026 State of Translation Automation is live today, and this year it's an interactive website, not a PDF. Choose your translation requirements, a target language and what to optimize for (quality, cost or latency), and see which solutions should work for you.
Card: The State of Translation Automation 2026, Intento annual industry report, previously The State of Machine Translation. 57 AI models and MT engines, 229 different solutions, 10 translation quality requirements, 100 target languages.

ALT Card: The State of Translation Automation 2026, Intento annual industry report, previously The State of Machine Translation. 57 AI models and MT engines, 229 different solutions, 10 translation quality requirements, 100 target languages.

1
24
After you specify your requirements, the next problem is making automatic translation meet them. For each language we use a solution complexity ladder: start with the simplest solution, often the fastest and cheapest, and add complexity and context until it meets them.
Card: solution complexity ladder from the 2026 benchmark, State of Translation Automation 2026, eight levels with the simplest at the bottom. 1, off-the-shelf NMT: an MT model as is. 2, adjusted NMT: native parameters such as glossary and formality. 3, off-the-shelf LLM: an LLM with a translation prompt. 4, adjusted LLM: RAG for dynamic business context. 5, fine-tuned models, NMT and LLM: not in the 2026 benchmark. 6, two-step solution: NMT or LLM, then automatic post-editing. 7, multi-step: translate, LQA gate, bounded repair. 8, multi-agent systems: agents that own different requirements cooperate. Each level adds context and usually costs more to build, run and maintain.

ALT Card: solution complexity ladder from the 2026 benchmark, State of Translation Automation 2026, eight levels with the simplest at the bottom. 1, off-the-shelf NMT: an MT model as is. 2, adjusted NMT: native parameters such as glossary and formality. 3, off-the-shelf LLM: an LLM with a translation prompt. 4, adjusted LLM: RAG for dynamic business context. 5, fine-tuned models, NMT and LLM: not in the 2026 benchmark. 6, two-step solution: NMT or LLM, then automatic post-editing. 7, multi-step: translate, LQA gate, bounded repair. 8, multi-agent systems: agents that own different requirements cooperate. Each level adds context and usually costs more to build, run and maintain.

1
39
Automating translation now means building a solution that meets your requirements, instead of choosing the "less wrong" off-the-shelf model. Building the harness around the models is iterative, every iteration needs evals, and one black-box 0–1 score can't do that job.
Card: 10 translation requirements from the 2026 benchmark, State of Translation Automation 2026, in three levels. Basic, for any company in this market: 1, general translation quality, MQM-based LQA minus the subjective requirements. Standard, for every company, with its own rules: 2, terminology, 3, tone of voice, 4, tags, 5, Unicode character integrity, 6, locale conventions, 7, HTML entities, 8, valid JSON, 9, spacing at tag boundaries. Advanced, for some companies and use cases: 10, gender neutrality.

ALT Card: 10 translation requirements from the 2026 benchmark, State of Translation Automation 2026, in three levels. Basic, for any company in this market: 1, general translation quality, MQM-based LQA minus the subjective requirements. Standard, for every company, with its own rules: 2, terminology, 3, tone of voice, 4, tags, 5, Unicode character integrity, 6, locale conventions, 7, HTML entities, 8, valid JSON, 9, spacing at tag boundaries. Advanced, for some companies and use cases: 10, gender neutrality.

1
17
In 2018 there were 18 MT vendors to choose a model from. Today 105 vendors of MT and LLMs supply the building blocks of a translation solution, and in the 2026 State of Translation Automation we benchmark 57 models from 37 of them. We publish the report on 13 October.
2018: choose the less wrong model. 2026: build a solution that meets your requirements. 18 MT vendors in 2018, 105 vendors of MT and LLMs in 2026.

ALT 2018: choose the less wrong model. 2026: build a solution that meets your requirements. 18 MT vendors in 2018, 105 vendors of MT and LLMs in 2026.

1
13
Our new Genesys app translates chat and email, and it can check a reply for offensive or harmful content before the support agent sends it. inten.to/language-hub-for-ge…
A customer writes in French, “Ma commande n’est pas arrivée”, and the support agent reads “My order hasn’t arrived”. The support agent replies “It ships tomorrow”, and the customer reads “Elle sera expédiée demain”. Between them, a card with the Genesys and Intento logos, and checks for terminology, tone of voice, content safety, and email plus chat.

ALT A customer writes in French, “Ma commande n’est pas arrivée”, and the support agent reads “My order hasn’t arrived”. The support agent replies “It ships tomorrow”, and the customer reads “Elle sera expédiée demain”. Between them, a card with the Genesys and Intento logos, and checks for terminology, tone of voice, content safety, and email plus chat.

1
4
58
1/ After the study of what people said about the localization job at LocWorld55, we went through what employers do, the open in-house positions and the people they appointed over the summer. The full analysis and the list of positions that are still open are in this thread.
Localization jobs in Q3 2026. AI labs are hiring, linguists in demand, 22 open director-level positions. 619 open positions, 308 recent appointments, 22 director-level seats. A donut of 927 new localization jobs by role type: translators and linguists 36%, management 20%, program and project 19%, engineering and i18n 7%, media production 6%, quality and LQA 4%, AI, MT and data 2%, other 6%.

ALT Localization jobs in Q3 2026. AI labs are hiring, linguists in demand, 22 open director-level positions. 619 open positions, 308 recent appointments, 22 director-level seats. A donut of 927 new localization jobs by role type: translators and linguists 36%, management 20%, program and project 19%, engineering and i18n 7%, media production 6%, quality and LQA 4%, AI, MT and data 2%, other 6%.

1
1
44
We've published Intento Translation Portal for ServiceNow in the ServiceNow Store. Your employees can translate text and documents in 650+ languages inside the Service Portal, with AI checks against company translation requirements. Launch details below.
Intento Translation Portal for ServiceNow: on-demand text and file translation inside the Service Portal, under company translation rules

ALT Intento Translation Portal for ServiceNow: on-demand text and file translation inside the Service Portal, under company translation rules

1
1
23
A translation can break the rules in your brief and still outscore the correct version on COMET. On locale formatting that happens 36 to 47% of the time, in all four languages. On gender neutrality in German and Spanish, more often than not.
Cover: A good score is not a good fit — a customizable framework for requirements-based LQA. Daria Sinitsyna, Lead AI Engineer at Intento, with the AMTA Québec City 2026 logo.

ALT Cover: A good score is not a good fit — a customizable framework for requirements-based LQA. Daria Sinitsyna, Lead AI Engineer at Intento, with the AMTA Québec City 2026 logo.

1
18
Two rooms at LocWorld55 applauded opposite claims about the localization job: the playbook is dead, and the manager isn't going anywhere. We counted 380 tasks and 107 recurring workflows across all 38 sessions. Both rooms were right. inten.to/blog/localization-r…
Report cover: The Localization Reshuffle, Measured. What 38 LocWorld sessions reveal about the job AI is actually building. 380 observed tasks, 43 canonical tasks, 107 recurring loops. A grid of task cards, each marked AI end to end, human plus AI, or stays human.

ALT Report cover: The Localization Reshuffle, Measured. What 38 LocWorld sessions reveal about the job AI is actually building. 380 observed tasks, 43 canonical tasks, 107 recurring loops. A grid of task cards, each marked AI end to end, human plus AI, or stays human.

1
33
Slide decks from #LocWorld55 put 21% of the localization job on AI end to end. The same speakers' talks and Q&A put it at 16%. Decks show the pipeline as designed. Talks show where people are still needed, and that's more tasks than the slides let on.
Slope chart: verdict shares counted from slide decks alone (166 tasks) against the full spoken record (380 tasks). AI end to end falls from 21% to 16%, human plus AI rises from 46% to 51%, stays human holds at 29%, wanted but not working yet holds at 4%.

ALT Slope chart: verdict shares counted from slide decks alone (166 tasks) against the full spoken record (380 tasks). AI end to end falls from 21% to 16%, human plus AI rises from 46% to 51%, stays human holds at 29%, wanted but not working yet holds at 4%.

1
19
In the 2026 State of Translation Automation we score a language only if 2 unrelated metrics agree on it: 2 embedding models, or 2 quality-estimation models. 7 of 100 fail both pairs, so we decided not to score them: Haitian Creole, Igbo, Maltese, Sesotho, Turkmen, Uyghur, Xhosa.
Card: before we score a language, 2 metric pairs must agree. Semantic similarity: LaBSE and Cohere embed v3. Quality estimation: COMET-Kiwi XL and MetricX-24. 93 languages: either pair agrees, we score them. 7: neither pair agrees, 'not reliably measurable'.

ALT Card: before we score a language, 2 metric pairs must agree. Semantic similarity: LaBSE and Cohere embed v3. Quality estimation: COMET-Kiwi XL and MetricX-24. 93 languages: either pair agrees, we score them. 7: neither pair agrees, 'not reliably measurable'.

1
1
94
A high COMET score doesn't mean the translation followed your termbase, your locale rules or your style guide: 46% of the 9,217 requirement violations we found score above 0.90. Daria Sinitsyna compares requirements-based LQA with COMET at AMTA 2026 tomorrow.
Cover for the AMTA 2026 talk: A good score is not a good fit - a customizable framework for requirements-based LQA. Talk, September 2. Daria Sinitsyna, Lead AI Engineer at Intento, with the AMTA Quebec City 2026 logo.

ALT Cover for the AMTA 2026 talk: A good score is not a good fit - a customizable framework for requirements-based LQA. Talk, September 2. Daria Sinitsyna, Lead AI Engineer at Intento, with the AMTA Quebec City 2026 logo.

1
1
25
Load more